Meta's settlement puts a number on age-estimation accuracy in one direction only

David Greene reads the Meta settlement provision by provision. Meta settled with 52 state attorneys general — every state except Florida, New Mexico and Texas, plus D.C., American Samoa, Guam, the Northern Mariana Islands and Puerto Rico. Within one year of the effective date Meta must apply one or more age assurance methods to every Facebook and Instagram user in the settling states, replacing the self-attested birthdate it has used since 2006. Users get sorted into 18+, 13–17, and under-13; the under-13 accounts get deleted.

The agreement writes accuracy floors into the text. Commercially available methods must hit a false-positive rate at or below 10% for ages 16–17 and 3% for ages 13–15 within a year. Meta’s own methods get a slower ramp — 14% and 7% at year one, 10% and 5% at year two — and the same standard applies to a Meta-built method only if it is not functionally identical to a commercial one. There is no ceiling on false negatives. An adult wrongly assessed as a minor gets an appeals process and nothing else. New users who decline age estimation are treated as Teen Users after fourteen days regardless of stated age, and Meta must run proactive monitoring of adult accounts to flag likely circumvention and push those users back through assurance.

Age assurance data carries a deletion requirement with a carve-out: “Retainable Data” — metadata about which method a user went through — may be held 90 days to identify circumvention patterns. Greene’s reading is that the settlement obliges Meta to collect, analyze and retain more about teenage users over the next ten years, not less. Meta has argued before that age verification belongs at the app-store layer with Apple and Google; the settlement has it consulting OS and app-store age signals rather than moving the obligation there.

Age is the textbook selective-disclosure attribute. An ISO 18013-5 mDL carries age_over_NN elements that answer the question without releasing the birthdate, the name or the document number, with the holder consenting to each presentation; the EUDI wallet is required to support the same check, and the zero-knowledge constructions being fitted to both remove the issuer signature as a correlation handle. The settlement reaches for estimation instead, and its retention carve-out is what that costs: ninety days of metadata on which method each user was put through, kept to find people working around it.

The limit is coverage. Nobody issues a credential to a twelve-year-old, so the population the settlement most wants to identify is the one a credential check reaches least. It does reach the adults being sorted downward, who hold credentials, and who are the ones the agreement leaves with an appeals process and nothing else.

False positives are capped, tested annually, and enforceable by 52 attorneys general. False negatives are neither capped nor measured.